DIFNet: Decentralized Information Filtering Fusion
Neural Network with Unknown Correlation in Sensor
Measurement Noises

Ruifeng Dong*a,b,c, Ming Wang*d, Ning Liua,b,c, Tong Guoa,b,c, Jiayi Kange,f,✉, Xiaojing Shend, Yao Maoa,b,c,
aState Key Laboratory of Optical Field Manipulation Science and Technology, Chengdu, 610209, Sichuan, China bInstitute of Optics and Electronics, Chinese Academy of Sciences, Chengdu, 610209, Sichuan, China cUniversity of Chinese Academy of Sciences, Beijing, 100049, Beijing, China dSchool of Mathematics, Sichuan University, Chengdu, 610000, Sichuan, China eBeijing Institute of Mathematical Sciences and Applications, Huairou district, 101408, Beijing, China fHetao Institute of Mathematics and Interdisciplinary Sciences (HIMIS), Shenzhen 518000, Guangdong, China
arXiv Supplementary Code

This video demonstrates how multiple ground-based observation devices exchange information through communication with each other, enabling target tracking in a distributed scenario.


Abstract

In recent years, decentralized sensor networks have garnered significant attention in the field of state estimation owing to enhanced robustness, scalability, and fault tolerance. Optimal fusion performance can be achieved under fully connected communication and known noise correlation structures. To mitigate communication overhead, the global state estimation problem is decomposed into local subproblems through structured observation model. This ensures that even when the communication network is not fully connected, each sensor can achieve locally optimal estimates of its observable state components. To address the degradation of fusion accuracy induced by unknown correlations in measurement noise, this paper proposes a data-driven method, termed Decentralized Information Fliter Neural Network ( DIFNet), to learn unknown noise correlations in data for discrete-time nonlinear state space models with cross-correlated measurement noises. Numerical simulations demonstrate that Decentralized IFNet achieves superior fusion performance compared to conventional filtering methods and exhibits robust characteristics in more complex scenarios, such as the presence of time-varying noise.


The DIFNet Framework

Decentralized IFNet Architecture

The fusion center refers to the processing facilities. (a) Centralized: the fusion center computes the global estimate directly from all sensor measurements $\vz$. (b) Hierarchical: low-level fusion centers compute local estimates $\hat{\vx}, \hat{\mP}$ from their own sensor measurements $\vz$, and a high-level fusion center combines these local estimates into a global estimate. (c) Decentralized: each fusion center computes a local estimate $\hat{\vx}, \hat{\mP}$ from its own measurements $\vz$, and then fuses it with estimates from neighboring centers to produce an individual fusion estimate.

Decentralized IFNet Architecture

Schematic diagram of the proposed DIFNet. The top part shows the generation of local estimates and error covariances by local sensors. The bottom part illustrates the communication and implementation flow at each fusion center."


  Experiments

Linear state space model

Image 1

RMSE of estimated position

Image 2

RMSE of estimated velocity

Nolinear state space model

Image 1

RMSE of estimated position

Image 2

RMSE of estimated velocity

Time-varying noise

Image 1

RMSE of estimated position (σ = 0.5)

Image 2

RMSE of estimated velocity (σ = 0.5)

* σ indicate measurement noise case